On non-standard configurations such as multi-card mining GPUs, llama.cpp is generally easier to adapt than vLLM.